\r曹玉珍¹，高晨阳¹，余 辉¹，张力新¹，王 江\r²\r
\r
AuthorsHTML:\r曹玉珍¹，高晨阳¹，余 辉¹，张力新¹，王 江\r²\r
\r
AuthorsListE:\rCao Yuzhen¹，Gao Chenyang¹，Yu Hui¹，Zhang Lixin¹，Wang Jiang\r²\r
\r
AuthorsHTMLE:\rCao Yuzhen¹，Gao Chenyang¹，Yu Hui¹，Zhang Lixin¹，Wang Jiang\r²\r
\r
Unit:\r1. 天津大学精密仪器与光电子工程学院，天津 300072；
2. 天津大学电气自动化与信息工程学院，天津 300072\r
\r
Unit_EngLish:\r1. School of Precision Instruments and Optoelectronics Engineering，Tianjin University，Tianjin 300072，China；
2. School of Electrical and Information Engineering，Tianjin University，Tianjin 300072，China\r
\r
Abstract_Chinese:\r\r传统癫痫发作通道选择方法需要提取特征，然后人工进行特征选择，最后基于所选特征训练分类器实现发作检测．为优化特征提取与选择过程，提出一种具有自学习特性，基于深度学习的癫痫脑电通道选择与发作自动检测组合模型．该方法利用卷积自编码器对癫痫脑电数据进行自适应特征提取，获得代表不同通道的特征子集；依据费舍尔准则筛选出特征子集与脑电通道；通过基于参数迁移的一维卷积神经网络实现癫痫发作脑电信号的检测．使用\rPhysioNet\r网站中的\rCHB-MIT\r数据库中\r8\r例有效数据量较为充足的病患脑电数据对组合模型进行有效性评价．对比该方法与基于方差、方差差异性和随机筛选方法得到的结果，在测试集上对癫痫发作检测的准确率、真阳性率、假阳性率的平均值分别达到了\r92.79\r%\r、\r93.07\r%\r、\r5.16\r%\r，均优于其他方法，且模型收敛速度所需的迭代次数平均仅为其他方法的\r10%\r．该方法在癫痫脑电发作检测效果和模型训练成本方面都有一定优势，且在进行脑电通道筛选时不需要手动提取特征，同时也可用于阿尔兹海默症等其他脑部疾病辅助诊断个性化检测模型的建立．\r\r
\r
Abstract_English:\rTraditional epileptic seizure channel selection methods require extraction of features，manual selection of features，and finally training the classifier based on selected features to achieve seizure detection. This study proposed a combined model with auto-learning characteristics based on deep learning to optimize feature extraction and selection process. The method used a convolutional autoencoder to perform adaptive feature extraction on epileptic electroencephalogram(EEG)data，and feature subsets representing different channels were obtained. Then，feature subsets and EEG channels were selected according to the Fisher criteria. Finally，the detection of seizure EEG signals was realized by one-dimensional convolutional neural network based on parameter transfer. Then，EEG data of 8 patients with sufficient data in the CHB-MIT database at PhysioNet were used for the effective evaluation of the combined model. Comparison of the results with the selection methods based on variance，variance difference，and random
demonstrates an average seizure detection accuracy of 92.79%，a true positive rate of 93.07%，and a false positive rate of 5.16% for this method on the test set. All results are superior to other methods，and the average number of it2020 erations required for model convergence is only 10% of the other methods. This method has certain advantages in terms of the detection effect of epileptic EEG and the training cost of the model and does not require the manual extraction of features when performing EEG channel selection. At the same time，it can be used to establish a personalized detection model for other brain disorders such as Alzheimer’s disease．\r
\r
Keyword_Chinese:癫痫发作检测；卷积自编码器；费舍尔准则；参数迁移\r

Keywords_English:epileptic seizure detection；convolutional autoencoder；Fisher criteria；parameter transfer\r


PDF全文下载地址:http://xbzrb.tju.edu.cn/#/digest?ArticleID=6443